Sweeteners and the gut microbiome: 10 trials, no signal

Sweeteners and the gut microbiome have been an open argument for a decade. Sucralose sits at the centre of it for a mechanical reason, stated in the review itself: roughly 85% of an ingested dose reaches the colon unmetabolised, so it meets the microbiota directly. That has made it the most-tested non-nutritive sweetener on this endpoint, and the one most often cited by brands and critics alike.

A systematic review published in September 2026 pooled the randomised human evidence: ten trials in healthy adults, seven of them with data usable for meta-analysis. Nothing reached statistical significance. The useful part of the paper is not that headline but the number underneath it: the authors ran a post hoc power analysis and published the result. It lands between 5.0% and 19.2%. That changes what the nulls mean, and it changes what a brand can write in a dossier.

What the review actually pooled

The search ran to May 2026 across PubMed, Scopus and Google Scholar, prospectively registered as PROSPERO CRD420261422049. It returned 4,864 records, 2,254 of them duplicates; 2,610 titles and abstracts were screened, 24 full texts assessed, and 10 randomised trials included. Seven contributed quantitative data. Individual trials enrolled between 8 and 61 participants, doses ranged from 48 to 780 mg of sucralose per day, interventions lasted one to ten weeks, and about 70% of the trials were run in North America. Pooling used a random-effects REML model with the Hartung–Knapp adjustment.

Six estimates came out of that. Shannon diversity was flat at a standardised mean difference of 0.00 (95% CI −1.16 to 1.15, three trials, I² 25.7%). Acetate came out at +1.45 (−2.37 to 5.28), propionate at +0.66 (−3.17 to 4.49), and butyrate at −0.42 (−2.46 to 1.62) from two trials that pointed in opposite directions, with I² of 88.1%. On the metabolic side, fasting blood glucose was −0.04 (−0.63 to 0.56, I² 69.6%) and HOMA-IR was −0.13 (−0.36 to 0.10, I² 0%). Composition itself was handled descriptively: three trials showed a numerically lower Firmicutes-to-Bacteroidetes ratio and one showed a higher one, but those ratios were reconstructed from published figures and were never statistically tested in the original trials. Beta diversity was reported by five trials and could not be pooled at all.

All effect sizes are standardised mean differences (Hedges g), sucralose minus control. Publication bias was not assessed for any outcome, because fewer than ten trials contributed to each.

Six null results at 5% power

Statistical power is the probability that an analysis would detect an effect of a given size if that effect is real. At 5%, an analysis is at the floor: it is no more likely to flag a true effect than a coin-flip study is to produce a false alarm. Four of the six estimates here sit between 5.0% and 6.3%. The best of them, HOMA-IR, reaches 19.2%. That is the number to carry out of this paper.

The intervals say the same thing in another unit. The acetate estimate spans 7.65 standardised units from one end to the other. A difference large enough to change anything in a human gut would fit comfortably inside that interval without being detected. So the correct reading of these six nulls is not that sucralose leaves the microbiota untouched. It is that these trials were too few, too small and too short to measure it. The authors say so themselves, and they also note that publication bias could not be ruled out, because no outcome had the ten trials a funnel plot would need.

That cuts both ways, and this is the part worth holding on to. A brand cannot cite this review as evidence that a sweetener is gentle on the gut. An activist campaign cannot cite it as evidence of harm. It is an absence of measurement, correctly reported.

The comparator decides what was measured

The review classifies the comparators used by the ten trials into two groups. Category A was sugar-free: water, capsules, citric acid solutions. Category B contained sugar: glucose, maltodextrin or dextrose. The authors are explicit that neither group is biologically inert, and that this heterogeneity was the single largest source of high risk-of-bias ratings, concentrated in domain 2, deviations from intended interventions. Three of the ten trials were rated high risk overall, four raised some concerns, three were low risk.

The consequence is structural rather than statistical. A trial that gives sucralose against glucose measures what happens when you take sugar out. A trial that gives sucralose against water measures what happens when you put a sweetener in. Those are two different questions, and a pooled estimate across both answers neither. Too few trials were available to split them in a subgroup analysis, so the review could only flag the problem. We have seen the same pattern decide the answer in blood pressure trials of sweeteners: the comparator is the intervention.

One sweetener is not a category

The premise that makes sucralose interesting here is molecule-specific. It is poorly metabolised, so most of a dose arrives in the colon intact and can interact with the resident bacteria. Nothing about that premise carries over to another sweetener, and the review measured no other sweetener. Headlines that convert this into a statement about artificial sweeteners in general are reporting something the paper did not test.

For steviol glycosides the closest randomised evidence on the same endpoint is older and smaller. A 2024 parallel-group, double-blind trial gave 59 healthy adults a daily beverage containing either 25% of the acceptable daily intake of stevia or 30 g of sucrose for four weeks, followed by a four-week washout, with shotgun sequencing of the microbiome and faecal short-chain fatty acids by UHPLC-MS/MS. At week four there were no significant differences between arms at phylum, family, genus or species level, and none in short-chain fatty acids; body mass index had risen by 0.3 kg/m² in the sucrose arm relative to stevia (p = 0.013), and no other cardiometabolic measure differed. The washout week-eight comparison was also null. That trial sits outside the 24-month window this series normally applies and its author list includes staff of an ingredient manufacturer, as recorded in PubMed; we cite it as the nearest randomised reading on steviol glycosides, not as a settled answer, and it reports no confidence intervals in its abstract, so it stays out of the chart above.

What it means for a formulator

The specification names a molecule, not a category. Sucralose is E 955. Steviol glycosides are four separate additives, E 960a from leaf extract, E 960b produced by fermentation, E 960c produced enzymatically and E 960d glucosylated, each with its own identity, purity criteria and label declaration. A microbiota paper about one of them is not evidence about the others, and a purchase specification that says stevia rather than naming the additive and demanding a glycoside profile with an acceptance criterion per component has not described what is being bought. That distinction is the whole basis of how we document SUCRAPRIME™.

Describe the substitution, not the sweetener. If the evidence changes direction when the comparator changes, so does the product story. Removing 12 g of sugar from a serving and adding a sweetener to an unsweetened base are different reformulations with different consequences for bulk, energy reduction and labelling. The technical file should say which one the product is.

The regulator’s own open question is about process. In its February 2026 re-evaluation of sucralose, EFSA confirmed the acceptable daily intake but could not conclude on a proposed extension of use into fine bakery wares, citing uncertainty about chlorinated compounds that might form across the range of baking conditions, and recommended the Commission consider revising the EU specifications for the additive. Whatever the sweetener, the matrix and the process belong in the stability file: an assay that reports a total at packing is not the same number as a profile at the end of shelf life. If you want the two documents that answer that for a glycoside system, ask us for the specification and a batch certificate.

What the review cannot support

The limits are the finding. Post hoc power between 5.0% and 19.2% means every null in this paper is uninformative rather than reassuring. Butyrate rests on two trials that disagree in direction, with I² of 88.1%, and the paper publishes two different 95% intervals for that single estimate, which is a reporting defect we flag rather than resolve. No outcome had enough trials for a publication bias assessment. Three of the ten trials carry a high overall risk of bias, driven by placebos that were not inert. The Firmicutes-to-Bacteroidetes ratios are descriptive values extracted from figures, not tested quantities, and beta diversity was dropped from the synthesis entirely.

Scope is narrow in two further ways the authors state plainly. Most interventions lasted under four weeks, so nothing here speaks to habitual, long-term consumption. And the participants were healthy adults with a normal body mass index, which is not the population a metabolic argument about sweeteners usually concerns. The honest summary is short: on this endpoint, in this population, over these durations, the randomised evidence for sucralose is thin and the question is open. For steviol glycosides it is thinner still, and that is a reason to specify the material carefully rather than to claim anything about it.